The AI CVE exploitation evidence story: Headlines are scarier than reality

This intelligence brief analyzes over 1,250 AI-related CVEs across 50 vendors and concludes that only 4.2% (33 CVEs) show measurable exploitation evidence in the wild; it provides per-tool profiles for the top ten AI products (e.g., LangFlow, LiteLLM, N8n, Gradio), highlights common exploitation types (RCE, sandbox escapes, auth/authorization gaps, prompt injection), lists scanner coverage gaps, and recommends mitigations including patching, sandboxing, authentication, network segmentation, and improved scanner identification.